53 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Score-Based Generative Modeling without Diffusion: Langevin MCMC All the Way
Genentech
Recording
Abstract
Score-based generation need not reverse a forward diffusion process. This talk presents an alternative based on Langevin MCMC at a fixed noise level, with effective noise falling as measurements accumulate. Suitable noise choices keep the sampling problems log-concave, and extensions connect the construction to discrete-data generation.
Topics
- Score-based generation, search World Wide
- forward diffusion process, broader topic: diffusion process
- Langevin MCMC, search World Wide
- fixed noise level, search World Wide
- effective noise, broader topic: noise
- measurements
- sampling problems, search World Wide
- discrete-data generation, search World Wide
- noise choices, search World Wide